#43333 cli(update -i): measure table cells in terminal columns, not bytes

How this was found. By a read of the code during the work on #43236. There is no user report. The npm registry does not allow a package name that is not ASCII, but a private registry can serve one, and a workspace name or a catalog name comes from the user's own package.json.

Before and after (NO_COLOR=1, piped stdout, the first number is Bun.stringWidth of the line):

before (1.4.3-canary.1+b52d51348)
 55 |   dependencies                  Current  Target  Latest
 54 |   ❯ □ dep                       1.0.0    1.0.0   2.0.0
 46 |     □ 日本語パッケージ  1.0.0    1.0.0   2.0.0

after
 47 |   dependencies          Current  Target  Latest
 46 |   ❯ □ dep               1.0.0    1.0.0   2.0.0
 46 |     □ 日本語パッケージ  1.0.0    1.0.0   2.0.0

On a 110-column pty, before: 工作区-文�… in the Workspace column (cut inside 档). After: 工作区-文档-😀 (14 columns, it fits in the 15-column limit).

With versions 1.0.0-canary.123 (16 columns) and a 15-column limit, before: 1.0.0-c…ary.123 1.0.0-canary.123 2.0.0-canary.4567. After: 1.0.0-c…ary.123 1.0.0-c…ary.123 2.0.0-c…ry.4567.

utf8_suffix_index_at_width. The suffix that fits is what remains after a prefix of at least total - max_width columns is dropped. utf8_index_at_width picks that prefix, so the two directions share one width rule, and the cut is never inside a code point or an ANSI escape. The forward function includes the zero-width code points that follow the prefix, so the suffix starts with a visible code point and a combining mark stays with its letter. When a 2-column code point straddles the cut, the first try drops one column too few. The next try drops that code point too. A code point is at most 2 columns wide, so two tries are enough, and a debug assertion checks the second one. The tests cover the exact cut (ASCII and decomposed é), the straddle (CJK) and a width of 0 (the Workspace column and the help text truncate at the end only).

Version columns. calculate_column_widths still sizes the Current, Target and Latest columns by byte length. A formatted semver version is ASCII (src/semver/Version.rs accepts only [0-9A-Za-z.+-] in a tag), so its byte length is its width. A truncated version holds a …, so the row renderer measures it with the width function.

Known limit. The UTF-8 width function adds the width of each code point. It differs from Bun.stringWidth on an emoji sequence (ZWJ, skin tone, keycap, VS16), as console.table and #43236 do. The tests do not pin the width of such a sequence. #41525 changes that in the shared function, and this table follows with no change here.

Tests. Verdaccio has no package with a name that is not ASCII, so the new tests use a manifest-only Bun.serve registry, bun install --lockfile-only and bun update -i --dry-run. The Workspace column needs a terminal wider than 100 columns, so that test runs bun update -i on a 110-column Bun.Terminal. It does not run on Windows: ConPTY redraws the screen with its own escape sequences, so the pty bytes are not the bytes that bun wrote.

Checked by hand with the debug build. Terminal widths of 50, 70, 80, 110, 140 and 170 columns, with and without colors. Names: Thai, decomposed é, a name that starts with combining marks, 50 zero-width spaces before the text, a tab and a bell, right-to-left text, flags, and mixed scripts. Every row puts the Current cell in the same column, and the table has no U+FFFD.

Same class, not changed here. These also pad or cut by byte length. Each is a different command with its own layout code and tests:

• bun outdated: #43236.
• src/runtime/cli/multi_run.rs (the label column of bun run --filter and --parallel).
• src/runtime/cli/build_command.rs (the output path column of bun build).
• src/runtime/cli/test_command.rs with src/sourcemap_jsc/CodeCoverage.rs (the file column of the coverage table).
• strip_ansi_len in src/runtime/cli/pm_diff_command.rs.
• buf_write in src/bun_core/Progress.rs cuts the progress line at the end of its buffer. With a long name of wide characters, the Resolving... line of bun install can end in a split character.